Cherrypicks for v1.8.x-aws

.github/workflows/distcheck.yaml

@@ -0,0 +1,58 @@
+name: PR CI
+on: [push, pull_request]
+env:
+  APT_PACKAGES: >-
+    build-essential
+    clang
+    gcc
+    git
+    libhwloc-dev
+    make
+jobs:
+  distcheck:
+    runs-on: ubuntu-22.04
+    strategy:
+      matrix:
+        cc:
+          - gcc
+          - clang
+      fail-fast: false
+    steps:
+      - name: Install Dependencies
+        run: |
+          sudo apt-get install -y ${{ env.APT_PACKAGES }}
+      - name: Install CUDA
+        run: |
+          sudo apt-get install -y nvidia-cuda-toolkit
+      - uses: actions/checkout@v4
+      - name: Build and Distribution Checks
+        run: |
+          set -x
+
+          # We're just doing distchecks, so it is fine if we
+          # just grab the latest master.
+          git clone --depth 1 https://github.com/ofiwg/libfabric.git
+          pushd libfabric
+          ./autogen.sh
+          ./configure --prefix=$PWD/install CC=${{ matrix.cc }}
+          make -j $(nproc)
+          make install
+          popd
+
+          # actions/checkout@v4 would drop the plugin source in $PWD,
+          # so go ahead and build it
+          ./autogen.sh
+          ./configure --with-libfabric=$PWD/libfabric/install --with-cuda=/usr/local/cuda/ CC=${{ matrix.cc }}
+          make -j $(nproc)
+
+          # Run Unit tests
+          make check
+
+          # Run dist tarball checks
+          make distcheck
+      - name: Upload build logs
+        if: failure()
+        uses: actions/upload-artifact@v4
+        with:
+          name: ${{ matrix.cc }}-config.log
+          path: config.log

src/nccl_ofi_api.c

@@ -24,14 +24,51 @@ ncclDebugLogger_t ofi_log_function = NULL;
 
 static ncclResult_t nccl_net_ofi_retval_translate(int retval)
 {
+	/*
+	 * This translates both ISO C errnos as well as libfabric errnos (up to
+	 * FI_ERRNO_OFFSET they are synonymous).
+	 */
 	switch (retval) {
 	case 0:
 		return ncclSuccess;
 		break;
 	case -EINVAL:
+		/*
+		 * Per ext-net docs, invalid arguments to plugin calls should
+		 * return ncclInternalError. Although a ncclInvalidArgument is defined,
+		 * it is suggested that ext-net plugins not pass these up and
+		 * leave NCCL API argument validation to NCCL.
+		 */
+		return ncclInternalError;
+		break;
+	case -EMSGSIZE:
+		/*
+		 * TODO: Per ext-net docs, this aligns with ncclInvalidUsage,
+		 * which is also defined in NCCL source, but for some reason
+		 * that error type is not available in err.h that we pull from
+		 * ext-net headers upstream. This needs to be fixed once the
+		 * ext-net header gets fixed to include ncclInvalidUsage.
+		 */
 		return ncclInvalidArgument;
 		break;
+	case -ECONNABORTED:
+	case -ECONNRESET:
+	case -ECONNREFUSED:
+	case -ENOTCONN:
+	case -EHOSTDOWN:
+	case -EHOSTUNREACH:
+		/*
+		 * Pass up ncclRemoteError (introduced in NCCL 2.13.4, but
+		 * missing in ext-net documentation) for any unrecoverable peer
+		 * reachability errors.
+		 */
+		return ncclRemoteError;
+		break;
 	default:
+		/*
+		 * Catch-all for other errors, including lifabric-specific error
+		 * codes.
+		 */
 		return ncclSystemError;
 		break;
 	}